Two things, both found by measuring rather than by reading the code.
The voice-activity filter inside the recogniser is off. It was costing
words: across a night of land-mobile captures it dropped 5-15% of what
the same model finds without it -- 491 against 507, 339 against 384,
263 against 310 -- because a single-word over between two
transmissions looks to a VAD exactly like the noise it exists to
remove, and on a scanner those short replies are the ones worth
having.
Turning it off has a cost, and the cost is that Whisper hands back
"You" for five seconds of hiss as confidently as it hands back a
sentence. So the whole capture is now asked once whether anything in
it rises above its own noise. Digital silence measures 0.0 dB of
contrast and hiss at any level 0.7, while the quietest real capture of
that night measures 8.9 and most measure 10-27; the bar sits at 3, an
order of magnitude clear of both. It can veto a capture but never trim
one, which is the whole difference between it and the filter it
replaces.
The second thing: callsigns like WQVF960 were being missed entirely.
The shape being matched was the amateur one -- prefix, district digit,
suffix -- and everything else the FCC licenses is written the other
way round, the letters first and then the digits. On the GMRS and
business channels that is most of what is said: nine callsigns across
five transcripts of one evening went by unrecognised, and now do not.
The shape is written as the three allocations that exist rather than
as "letters then digits", which claims KN95, WD40 and KC135. Its
letters are checked against the word list even when they arrive as a
single token, which the amateur shape does not need -- no English word
has a digit in the middle of it, but "west 120" and "word 100" fit
this one exactly.
Lookups now fall back to hamdb.org when callook has nothing. Not a
spare copy: callook holds United States amateur licences only, so
DL1ABC and VE3ABC are INVALID there and resolve perfectly well from
the other. And a GMRS callsign is not looked up at all -- every
database reachable without an account is an amateur register, so
reporting WQVF960 as "unlisted" would blame the callsign for the
absence of a source.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
Under the transcript, headed DETECTED CALLSIGNS:, every callsign heard in
it with the name and location on its licence.
Finding them is not one regular expression over the text as written. A
speech recogniser is poor at callsigns -- they are not words, they are
said one character at a time -- so it breaks them wherever the speaker
paused and writes the phonetic alphabet down verbatim. The recording that
prompted this has "Alright, KU 0W" in it, with a space; spelled out it
would have been "kilo uniform zero whiskey". All three forms read back to
KU0W.
Not inventing them matters more. A run of words is accepted only when
none of its parts is an ordinary English word: "or 3. Can you open 4" and
"CC1 boy", both from real transcripts here, fit the shape once the
punctuation is gone and are not callsigns. A single token said in one
breath is still trusted, because W1BOY is a perfectly good callsign, and a
lone "a" or "i" cannot start a join or "a B4U player" becomes AB4U.
Across the 126 transcripts in the recordings directory that turns three
candidates into the one that was actually said.
Lookups use the FCC's own licence data at callook.info -- no account, no
key, the callsign the only thing sent. They never delay the display: the
entry reads "looking up" and fills itself in, and results are cached under
~/.cache so a net recorded night after night is looked up once.
--no-lookup contacts nothing and still describes a callsign from its own
structure, the ITU prefix giving the country and the digit the US
district, which is also all there is to say for callsigns outside the US.
--callsigns prints everyone who identified themselves and where they were
heard.
Also asked: are transcripts appended to, or overwritten, when another
transmission arrives on the same frequency? Neither could be shown from
reading the code alone, so there are now three tests that run real scans
and look at the files. By default each transmission has a transcript of
its own -- the timestamp is in the name, so two overs cannot land on one
file. With --combine there is one recording per frequency and therefore
one transcript, opened for append with the time of each over; a second
scan into the same directory adds to it rather than starting it over,
which is the case the last of the three tests covers.
The browser and callsign tests refuse to reach the network at all. One
test did, quietly, and passed -- visible only because the assertion it
failed printed a real operator's address.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016PsWPTweCT6pwxKngvVxcg
No speech recogniser is in Debian, so installing bandsaunter from a .deb
left transcription to a manual pip step on every machine. A repository of
one's own is not bound by archive policy, so build-repo.sh now packages
faster-whisper and the base.en model alongside the application:
bandsaunter the application (Architecture: all)
bandsaunter-transcribe faster-whisper, vendored (amd64)
bandsaunter-model-base-en the model, so nothing reaches the network
The wheels land in /usr/lib/bandsaunter/vendor rather than dist-packages,
and transcribe.py appends that directory to sys.path -- appends, so an
apt-managed numpy or PyYAML still wins and the vendor copy only fills the
gap. Duplicates of what Debian already ships are stripped from the tree.
resolve_model() turns a bare "base.en" into the packaged copy when one is
installed, and leaves it alone to be downloaded when none is.
The app package recommends the other two, so "apt install bandsaunter"
brings the lot and --no-install-recommends still gets just the scanner.
Its postinst explains how to add a recogniser only when there genuinely
is not one -- including the case where apt has already unpacked the
recogniser package but not yet configured it.
Verified with the source tree hidden and no home directory: the packaged
CLI runs, and a real recording transcribes offline from the vendored
engine and packaged model while numpy still resolves to the system one.
apt itself resolves the repository over HTTP and plans all three.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Sweeps any set of frequency ranges, records what it finds, and works out
what kind of signal it was.
- Frequency ranges entered by hand or picked from a 135-entry US band plan,
including whole-band and all-CW sweeps that resolve the demodulator per
segment.
- Detection calibrated against the peak-hold detector's own noise statistics,
so the threshold means real margin over static rather than over the floor.
- A content gate: captures are kept only if they carry voice, decodable CW,
or an identified digital keying scheme. Speech is recognised by a pitch
track that drifts, which static cannot imitate.
- Identification of NFM/WFM/AM/SSB, CW with Morse decoded to text, P25, DMR,
NXDN, D-STAR, POCSAG, FLEX, ACARS, AIS, APRS, n-FSK and n-PSK.
- Gapless streaming capture, with the signal path fast enough to keep up in
real time, so recordings play back at the right speed.
- Optional one-file-per-frequency recording with spoken timestamps, and
speech-to-text transcription.
- Menus and command line generated from one settings table, so neither can
offer something the other cannot; settings persist in ~/.config.
367 tests, run against synthetic signals, a built-in receiver simulator, and
real hardware.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
